I think you are talking about Grice's maxims
In the early days of natural language understanding (in the GOFAI era), there was a principle of "minimum surprise". An ambiguity would be resolved by choosing the least surprising of multiple possible interpretations, which one could also describe as the least interesting, or the most boring, or least informative. Which are different ways of saying the most likely.
"Water" has a range of senses, but for any particular use of the word, some particular referent will be intended. The most likely is, by definition, the one to guess.
David Chapman is wringing a whole book out of this, with no end in sight, although I see nothing in the eponymous dialogue that is in any way a problem for ordinary rationality.
This is basically Davidson's "principle of charity".
The above might be a communication failure; A might have been looking for some water to drink, and you can't drink the water in the cells of an eggplant, so it is not water that is suitable for the purposes of A's quest. We might pin the failure to communicate on B; the contents of the cells of the eggplant is not usually considered water.
But water has a core essence, a property that gives it all its other properties: H2O. (Or H2O in its liquid state, or similar; it doesn't really matter as the eggplant also contains H2O in its liquid state.) It literally is the central node in a network of all the properties of water. And there's similar cases where you would consider the water in the cells of the eggplant to be water; for instance you might say that the eggplant is mostly made up of water if you want a rough guess of its density.
I would tend to say that this shows that the meanings of words is contextual; it depends on the situation. But in what way, exactly, does it depend on the situation? Here's a proposal:
You guess a variety of facts that the people communicating might be interested in knowing about, as well as a variety of observables that are available, as well as people's epistemic state with respect to those observables, and then you select the meaning of your words so that they maximize the correlations with these, within the range of ambiguity that the words can refer to.
The idea being that this maximizes opportunities for relevant communication, as well as for deduction. This proposal is vague and not necessarily well thought-out; I would be interested in general discussion, pointers to other work, examples or counterexamples, etc..